Zol AI Lab
← All projects
Research · since October 2026research

Intelligent Tolling

A toll road, rebuilt with AI.

Seven phases, from live traffic on camera to a toll you can trust: plate reading and vehicle classification for Australian motorways, built phase by phase on our own NVIDIA GB10.

Toll road operatorsRoadside systemsTraffic data
Phase 07 · Verify: the interactive story →
96–100%character accuracy, before any Australian training
14 / 19legible test plates read reliably
1wrong-owner bill caught on test footage

Seven phases, one toll

Each phase gets its own page as it ships.

01Captureplanned

Real traffic on a dedicated ultra-HD, high-frame-rate camera, so every plate is sharp in many frames.

02Detectin progress

Every vehicle, plate and wheel found and tracked frame by frame.

03Readbaseline

Each plate read in every frame, then voted character by character into one answer.

04Learnplanned

Detector and reader retrained on Australian plates from every state, real and synthetic.

05Classifyplanned

Axles and axle groups counted and turned into the Austroads 12-class code.

06Measureplanned

Speed from a calibrated road plane, plus length, width and height in metres.

07Verifylive

Unsure plates checked against registration data and the vehicle on camera, before anyone is charged.

Read the live phasePhase 07 · Verify: the interactive storyReal footage, real reads: how the system settles a plate the camera isn't sure about, before anyone is charged.Open the story →

How it works

From input to finished result.

1
Motorway video4K and 1080p roadside footage
2
Find platesTiled detection so small plates in 4K survive
3
Read and trackOCR per character, tracked across frames, then voted
4
Check formatAustralian state formats and signwriting filter
5
VerifyRegistration lookup plus the vehicle on camera, described by a local vision-language model

What it does

The parts that make Intelligent Tolling useful.

Plate reading

A small plate detector and per-character OCR, with every read of a vehicle combined into one confident answer.

YOLOv9 plate detector · fast-plate-ocr · ByteTrack

Verifier

For uncertain reads: lists the likely spellings, looks each up, and compares the registered make and body with what the camera sees.

RF-DETR · Qwen3.6 vision-language · SigLIP 2

Decisions

VERIFIED, LIKELY, REVIEW or UNVERIFIED, so people only check the hard cases.

VERIFIED · LIKELY · REVIEW · UNVERIFIED

Privacy by design

Public roadside footage used privately for testing, no plates published, and open licences (Apache or MIT) preferred for every model.

Testing only · no plates shown

Models and tools

What Intelligent Tolling runs on.

Names and logos belong to their owners and identify the models and tools we use; no endorsement is implied.

Built so far

Engineering that is already in place.

  • Pretrained models only so far: the baseline the Australian models must beat
  • Vehicle found with RF-DETR segmentation; colour from the vehicle's own pixels
  • A local vision-language model describes the vehicle without seeing the plate
  • Test footage is public video used privately; no plates are published

Where it's going

The next steps, in order.

  1. Australian modelsTrain on synthetic and real Australian plates, add a state output; target 95%+ OCR on held-out crops
  2. Vehicle classesAxle counting and Austroads classes for trucks and trailers
  3. Speed and sizeCalibrated speed from the road plane, then vehicle dimensions

Lab notes

Progress on Intelligent Tolling.

Intelligent Tolling: checking uncertain plate reads

When a camera is unsure whether a character is a 1, an I or a T, the wrong person can be billed. Our verifier lists the likely spellings, looks each one up in the registration record, and compares the registered vehicle with what the camera actually sees: body type, make and badge, described by a local vision-language model that never sees the plate.

research · computer vision

Designed and built at Zol AI Lab by Moe Eldouma, founder. See his role, skills and other work on eldouma.dev.